# Reproduction and evidence

Verify `manifest.json` SHA-256 entries before use. The input-generalization run uses the pinned original spiking baseline under `runs/2026-10-26/2026-10-26_fly-baseline_v01`, plus the prior `runs/2026-10-28/t4-specificity-v01` membership and pathway results. The visual run uses that same spiking baseline, `runs/2026-10-28/visual-steering-v01` saved visual responses, bridge membership, pathway metadata and prior baseline outputs, and `runs/2026-10-29/visual-rectify-analysis-v02/comparison.npz`. The public bundle preserves the executed programs and saved results, but is not a self-contained copy of the original models. Reconstruct the expected directory tree and pinned dependencies from the earlier public experiment manifests before running; execute into a fresh directory to avoid overwriting records.

Each protocol and runner was hashed in `frozen-sha256.txt` before its own experiment. The visual rectification v01 analysis failed its old-encoder reconstruction assertion because it omitted the previous 1e-6 threshold. The v02 analysis applies that threshold to both old and new arrays; the failed v01 files and redacted traceback are preserved. The visual trial used the same random seeds as the earlier bridge, not held-out seeds. Uniform random draws were matched across old/new, but rates changed and therefore actual events differed: see `driver-comparison.json`. Within each new pattern, intervention/restoration conditions share the exact driver events.

The input audit rereads saved trial files and counts. The visual audit compares saved driver and ordered network-spike events and full-weight hashes for the restoration controls. Neither audit reruns neural dynamics or verifies unrecorded membrane trajectories. Paths in public logs were redacted; original logs remain in the private Linux run directories.